LEMMS: Label Estimation of Multi-feature Movie Segments

Bartolomeo Vacchetti, Dawit Mureja Argaw, Tania Cequtelli

Abstract: In the last few years, there has been an increasing amount of methods and algorithms that approach and automate different video and image editing tasks. A task that so far has not been investigated too much in depth is the analysis of video editing patterns. In this work, we present LEMMS (Label Estimation of Multi-feature Movie Segments), a methodology to analyze and label 30-second-long movie editing patterns based on the following editing features: shot size, shot subject, editing pace, and editing trend. LEMMS can identify more or less fine-grained editing classes using a multi-clustering approach. To evaluate the robustness of LEMMS in assigning correct labels the performance of an LSTM classifier is analyzed. For our study, we extracted 24 363 segments of movie scenes from the AVE [1] dataset. The performance of LEMMS in semi-automatic label identification for 30-second long movie segments is accurate, as the proposed approach has an overall accuracy of 92.8% for 50 classes.